GPT 6 Astra Built an Interactive 3D Apartment! (Higgsfield + Godot)

A walkable apartment before you move in

This is a prototype called Homemade 3D. It uses OpenAI's GPT-6 Astra, Godot, Blender, and Higgsfield to build an interactive 3D apartment. You can walk around, switch to third person, change furniture, and even sit on the sofa to check the TV sightline. The point isn't a pretty render. It's answering practical questions like: Will a wider sofa make the walkway too tight? Does the balcony light make the room feel open? And can two people actually move around the coffee table without squeezing past each other?

The version shown here has three default templates. You can pick a Japandi layout, which comes with coordinated furniture and finishes already arranged. From there, you move pieces using a top-down floor plan, then step back into first-person view to see how the change feels. The measuring tool is the most useful part. Instead of going back to a real apartment with a tape measure every time you want to test a different arrangement, you measure the room and the furniture in the app. The catch is the model must match the real measurements. Once it's set up, comparing options is fast.

Together mode and wall notes

The app is aimed at couples, families, or friends moving into a new place together. Together mode lets another person join the same apartment. You can suggest moving a sofa while they look at it from the other side of the room. You're both in the same space with your own viewpoints, which makes it easier to explain what feels awkward.

You can also leave wall notes, which are comments attached to a location. It's like working on the same Google document, but for a room. If you're not online at the same time, you can still leave opinions and come back to them later. There's also an NPC feature. You can add a few simulated people moving around to see where furniture blocks the route. A room can look spacious with one person, but a crowd tells a different story.

After the layout works, you can change the wall color, floor finish, and lighting. The night preview shows a warm glow from the floor lamp. The daylight preview shows the balcony as the main light source. It's an illustrative simulation, not an exact render, but it gives you a clear brief to communicate with a designer or contractor.

How the pieces fit together

Astra is the coordinator. Inside ChatGPT, Astra uses an MCP connection to Higgsfield, which lets it request images and video during the conversation. Godot runs the interactive scene: movement, furniture, and controls. Blender prepares the assets. For example, the linen wallpaper was authored in Blender with separate information for color, roughness, and fine surface detail. Those were exported as texture maps, a process called baking, so Godot can use them. There are also Blender scripts for character assets, similar to a 2K MyPlayer mode.

The workflow is interactive first, then presentation. You explore, make decisions, and get a space you can inspect. Then Astra and Higgsfield help show those decisions to others. That comparison, between the playable room and the rendered treatment, tells you whether the design ideas survived. The creator mentioned a few personal details: 80% of his own furnishings are IKEA, and the catalog pulls real product data including dimensions. That last point matters. You can plan around pieces you might actually buy, not generic placeholder furniture.

The missing feature is obvious. Uploading photos of a real apartment to generate the starting model would remove the need for three templates. The creator calls this the next feature he'd love to build. Until then, Homemade 3D is useful for testing ideas, but the manual setup and measurement matching are on you.